Anthropic Викторина

Ответы и пояснения появляются по мере прохождения теста.

1. Команда оценивает Claude в claude.ai и через API разработчика. Чего следует избегать?
   • That the application, API and hosted integrations necessarily expose identical features and controls
   • That different deployment surfaces can have different settings
   • That it should test the surface used by its own application
   • That it needs the exact endpoint for its evaluation
2. Команда считывает системную карту Claude для выпуска модели. Что устанавливают эти доказательства?
   • That every application using the model will have the same behavior
   • Evidence and limitations for that release, not a guarantee about every downstream application
   • That the model can never make errors outside tested tasks
   • That the team no longer needs application-specific evaluation
3. Приложение на основе Claude получает запрос инструмента на стороне клиента. Кто выполняет эту операцию?
   • The model by itself
   • The user who wrote the prompt, regardless of the application
   • The application handling the client-side tool call
   • The system card for the model
4. Помощник с помощью инструментов сообщает, что обновил запись. Что должно проверять приложение?
   • Whether the narration sounds confident
   • Whether the model card lists tool use
   • Whether the prompt named the record
   • Whether the operation completed and the resulting state is correct
5. Команда сравнивает Claude через облачную платформу с ее прямым API. Что входит в протокол оценки?
   • The exact endpoint, model version and relevant configuration
   • Only the Claude family name
   • Only the launch date of the application
   • The user’s favorite answer
6. Why should a developer create task-specific tests instead of relying only on broad quality claims?
   • A model card replaces testing for every application
   • Measured behavior depends on the task and evaluation conditions
   • A system card guarantees high quality in all contexts
   • Testing makes platform differences disappear
7. A retrieval tool fails while answering a policy question. What behavior should an application test for?
   • A confident answer invented from general context
   • A claim that the tool succeeded without checking
   • A clear report that the evidence is missing rather than an invented policy
   • A change to the policy text by the model
8. Before changing a Claude model version in production, what should a team plan?
   • Assume a stable API means output behavior cannot change
   • Delete the evaluation set to avoid conflicting results
   • Skip rollback planning if a new release is available
   • Repeat the relevant evaluation and maintain a deprecation or rollback plan